Data set normalization in torchvision_tutorial.html

Hello,
I have been implementing this tutorial :
https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html
The input images are not normalized to 0 mean and 1 variance.
Is this “normal” ? perhaps an oversight ?
Best regards

It’s a good question to ask.
The answer is that one has to consider whether inputs are expected to be normalized by the model - a very natural alternative is to consider normalization part of the model.

In this case we can refer to the TorchVision documentation:

The input to the model is expected to be a list of tensors, each of shape [C, H, W], one for each image, and should be in 0-1 range. Different images can have different sizes.

But again, you are absolutely right that this is something to check (and maybe it would be worth mentioning in the tutorial, too).

Best regards

Thomas